Terminal-Bench-RL: Training long-horizon terminal agents with RL

In plain words

Terminal-Bench-RL is an infrastructure for training reinforcement learning agents to perform complex terminal and coding tasks over long horizons. It uses Docker-isolated training environments and a hybrid reward system combining unit test verification with LLM-based behavioral evaluation. The system scales from small GPU setups to large clusters and is designed for researchers and developers building autonomous agents capable of executing multi-step programming tasks. A 32-billion-parameter agent trained with this system achieved competitive results on the Terminal-Bench leaderboard without additional training.

written from the facts on this page · September 2026

From the sources

GRPO training code which scales to 32xH100s for long horizon terminal/coding tasks. Base agent is now the top Qwen3 agent on Stanford's TerminalBench leaderboard. - Danau5tin/terminal-bench-rl

In the maker’s words, at launch

After training calculator agent via RL, I really wanted to go bigger! So I built RL infrastructure for training long-horizon terminal/coding agents that scales from 2x A100s to 32x H100s (~$1M worth of compute!) Without any training, my 32B agent hit #19 on Terminal-Bench leaderboard, beating Stanford's…

• 
  I RL-trained an agent that trains models with RL (for ~$1.3k)Jul 2026 · github.com · ▲107

  RL-training an AI agent to RL-train AI agents. Contribute to Danau5tin/ai-trains-ai development by creating an account on GitHub.

• 
  Agent 37Mar 2026 · agent37.com · ▲401

  Hey HN, I noticed there are 1,000+ MCP skills on GitHub with zero way for creators to charge for them. So I built Agent37. Selling Claude skills today means asking customers to download your skill, set up Claude Code on their laptops, configure MCP servers, and pray it works. No trial, no updates, often you're just handing over the source. Agent37 lets creators upload a skill, share a link, and anyone can try it instantly (no Claude account needed). Built-in Stripe, creators keep 80%. Looking for feedback on the approach. Would you use something like this?

• 
  SupersetFeb 2026 · superset.sh · ▲541

  Hey HN, we’re Avi, Kiet, and Satya. We’re building Superset, an open-source terminal made for managing a bunch of coding agents (Claude Code, Codex, etc) in parallel. - Superset makes it easy to spin up git worktrees and automatically setup your environment - Agents and terminal tabs are isolated to worktrees, preventing conflicts - Built-in hooks [0] to notify when your coding agents are done/needs attention, - A diff viewer to review the changes and make PRs quickly We’re three engineers who’ve built and maintained large codebases, and kept wanting to work on as many features in…

• 
  OSS Agent I built topped the TerminalBench on Gemini-3-flash-previewApr 2026 · github.com · ▲393

  Scored 65.2% vs google's official 47.8%, and the existing top closed source model Junie CLI's 64.3%. Since there are a lot of reports of deliberate cheating on TerminalBench 2.0 lately (https://debugml.github.io/cheating-agents/), I would like to also clarify a few things 1. Absolutely no {agents/skills}.md files were inserted at any point. No cheating mechanisms whatsoever 2. The cli agent was run in leaderboard compliant way (no modification of resources or timeouts) 3. The full terminal bench run was done using the fully open source version of the agent, no…

• 
  Spotlight by BackplanesJun 2026 · backplanes.com · ▲394

  Hola HN! Long time lurker, sometimes commentor, first time poster here. I’ve been working alongside my two co-founders and a few colleagues on a project I’m excited (and a little nervous) to share with you all! Like many of us, I’ve lived a tortured existence with AI coding (is it vibes?) over the past few months - I think November was a big moment with this. But, one thing I’ve noticed after building orchestration layer after orchestration layer is that the thing I always came back to was “what the hell is Claude Code actually doing?” Perhaps it’s because of the time Claude Code got…

• 
  20+ Claude Code agents coordinating on real work (open source)Feb 2026 · github.com · ▲53

  Single-agent LLMs suck at long-running complex tasks. We’ve open-sourced a multi-agent orchestrator that we’ve been using to handle long-running LLM tasks. We found that single LLM agents tend to stall, loop, or generate non-compiling code, so we built a harness for agents to coordinate over shared context while work is in progress. How it works: 1. Orchestrator agent that manages task decomposition 2. Sub-agents for parallel work 3. Subscriptions to task state and progress 4. Real-time sharing of intermediate discoveries between agents We tested this on a Putnam-level math problem, but the…

• 
  Spira Maxima▲578

  Video model that turns script into viral social media videos

  AI · 18d ago · spira.ai

• 
  Rill Browser▲540

  The browser where Claude Code and Codex work beside you

  AI · 18d ago · rill.love

• 
  Germany's new sovereign AI model Kolibri▲424

  Kolibri is Aleph Alpha

  AI · 20d ago · tej.as

• 
  Giving Opus 5.5 a simulated paint canvas▲392

  Paintings by AI models: programs run through a simulation of oil paint.

  AI · 22d ago · stillwet.art

• 
  Beam.win▲378

  Run AI agents on one Mac and use its windows on another

  AI · 5d ago · beam.win

• 
  FastRouter.ai▲364

  Route requests to the right LLM for cost, latency & quality

  AI · 23d ago · fastrouter.ai

• 
  Clueso▲1,300

  Create stunning product videos in minutes with AI

  AI · 2025 · clueso.io

• 
  Trickle - Magic Canvas ▲1,064

  Hey HN! Introducing Trickle Magic Canvas: the world’s first Agentic Canvas, where you can visually co-create apps and websites with AI, from idea to production. Why is a canvas better? Most AI tools today rely on linear chats or text prompts. We believe that limits creativity and clarity. Canvas changes the game. 1. Every version & page at a glance. Each page and sub-page lives on a timeline. You can quickly browse, edit, and preview progress like a living, visual document — no more digging through chat logs or scattered states. 2. The canvas is the context. Unlike linear chats, you can…

  AI · 2025 · trickle.so

• TB
  Tinder but it's only pictures of my wife and I can only swipe right▲1,030

  Life & fun · 2025 · trytender.app

• 
  Ten years of running every day, visualized▲958

  Today marks ten years, 3653 consecutive days, of running at least one mile every day under the USRSA rules [1]. To celebrate, I built an interactive dashboard that turns a decade of GPX files into charts you can explore. Running has truly changed my life: I've made lifelong friends, explored beautiful places, and more importantly invested into my own health and fitness, which I'm starting to see the positive benefits as I get older. The stack is pretty simple: a NextJS app, with a Postgres database to keep all my running data, and all the stats are pre-computed and cached in Redis, so I…

  Dev tools · 2025 · nodaysoff.run

• 
  Draw a fish and watch it swim with the others▲935

  Made this website as an exercise in vibe-coding and GCP. It was posted about a few times around the internet, on sites like Morning Brew (https://www.morningbrew.com/issues/business-buzzkill), MetaFilter (https://www.metafilter.com/209703/Draw-A-Fish), boingboing.net, etc. I think it's cute! I built a basic CNN trained against penises and swastikas, and then anything that doesn't hit the 63% confidence score gets sent to a mod queue, a vibe-coded fish-tinder…

  AI · 2025 · drawafish.com

• 
  Use Their ID – Use your local UK MP’s ID for the Online Safety Act▲862

  Hi HN - I made a site that takes a UK postcode, grabs the local MP's information and generates an AI mockup of what their ID might look like. It's a small, silly protest at the stupidity of the Online Safety Act that just came into force. edit - My open AI credits got hugged to death, please use a known postcode (like one from Kier Starmer's constituency, WC2B6NH) in the meantime.

  AI · 2025 · use-their-id.com